Update Network Parameters After Code Generation

This example shows how to update learnable and state parameters of deep learning networks without regenerating code for the network. You can update the network parameters for SeriesNetwork, DAGNetwork and dlnetwork.

Parameter update supports MEX and standalone code generation for the NVIDIA^® CUDA^® deep neural network library (cuDNN) and the NVIDIA TensorRT™ high performance inference libraries.

Create an Entry-Point Function

Write an entry-point function in MATLAB^® that:
1. Uses the coder.loadDeepLearningNetwork function to load a deep learning model. For more information, see Load Pretrained Networks for Code Generation.
2. Calls predict (Deep Learning Toolbox) to predict the responses.

For example:

function out = mLayer(in, matFile)

myNet = coder.loadDeepLearningNetwork(coder.const(matFile));

out = myNet.predict(in);

Create a Network

The network used in this example requires input images of size 4-by-5-by-3. Create sample network inputs of the same size format as the network inputs.

inputSize = [4 5 3];
im = dlarray(rand(inputSize, 'single'), 'SSCB');

Define the network architecture.

outSize = 6;
layers = [
    imageInputLayer(inputSize,'Name','input','Normalization','none')
    convolution2dLayer([3 3], 5, 'Name', 'conv-1')
    batchNormalizationLayer('Name', 'batchNorm')
    reluLayer('Name','relu1')
    transposedConv2dLayer([2 2], 5, 'Name', 'transconv')
    convolution2dLayer([2 2], 5, 'Name', 'conv2')
    reluLayer('Name','relu2')
    fullyConnectedLayer(outSize, 'Name', 'fc3')
    ];

Create an initialized dlnetwork object from the layer graph.

rng(0);
dlnet1 = dlnetwork(layers);
save('trainedNet.mat', 'dlnet1');

Code Generation by Using `codegen`

To configure build settings such as output file name, location, and type, you create coder configuration objects. To create the objects, use the coder.gpuConfig function.

To specify code generation parameters for cuDNN, set the DeepLearningConfig property to a coder.CuDNNConfig object that you create by using coder.DeepLearningConfig.

cfg = coder.gpuConfig('mex');
cfg.TargetLang = 'C++';
cfg.DeepLearningConfig = coder.DeepLearningConfig('TargetLibrary', 'cudnn')
cfg.DeepLearningConfig.AutoTuning = true;
cfg.DeepLearningConfig.DataType = 'fp32';

Specify the inputs.

cnnMatFile = fullfile(pwd, 'trainedNet.mat');
inputArgs = {im, coder.Constant(cnnMatFile)};

Run the codegen command. The codegen command generates CUDA code from the mLayers.m MATLAB entry-point function.
```
codegen -config cfg mLayer -args inputArgs -report
```

Run the Generated MEX

Call predict on the input image and compare the results with MATLAB.

out = mLayer_mex(im,cnnMatFile)
out_MATLAB = mLayer(im,cnnMatFile)

out1 = 

  6(C) x 1(B) single dlarray

   -0.0064
   -0.1422
   -0.0897
    0.2223
    0.0329
    0.0365


out_MATLAB = 

  6(C) x 1(B) single dlarray

   -0.0064
   -0.1422
   -0.0897
    0.2223
    0.0329
    0.0365

Update Network with Different Learnable Parameters

Re-initialize dlnetwork to update learnables to different values.

rng(10);
dlnet2 = dlnetwork(layers);
save('trainedNet.mat', 'dlnet2');

Use the coder.regenerateDeepLearningParameters function to regenerate the bias files based on the new learnables and states of the network.

The first input to the coder.regenerateDeepLearningParameters function is a SeriesNetwork, DAGNetwork or dlnetwork object. The second argument is the path to the network parameter information file emitted during code generation. You can optionally specify the NetworkName=MYNET name-value pair to specify the name of the C++ class for the network in the generated code.

codegenDir = fullfile(pwd, 'codegen/mex/mLayer');
networkFileNames = (coder.regenerateDeepLearningParameters(dlnet2, codegenDir))'

The coder.regenerateDeepLearningParameters function returns a cell-array of files containing network learnables and states.

networkFileNames = 

  8×1 cell array

    {'cnn_trainedNet0_0_conv-1_b.bin'   }
    {'cnn_trainedNet0_0_conv-1_w.bin'   }
    {'cnn_trainedNet0_0_conv2_b.bin'    }
    {'cnn_trainedNet0_0_conv2_w.bin'    }
    {'cnn_trainedNet0_0_fc3_b.bin'      }
    {'cnn_trainedNet0_0_fc3_w.bin'      }
    {'cnn_trainedNet0_0_transconv_b.bin'}
    {'cnn_trainedNet0_0_transconv_w.bin'}

Note

For MEX workflows, when the generated MEX and the associated codegen folder is moved from one location to another, coder.regenerateDeepLearningParameters cannot regenerate files containing network learnables and states parameters in the new location. Set the 'OverrideParameterFiles' parameter of coder.regenerateDeepLearningParameters to true to allow the coder.regenerateDeepLearningParameters function to regenerate files containing network learnables and states parameters in the original codegen location.

For standalone workflows, coder.regenerateDeepLearningParameters can regenerate files containing network learnables and states parameters in the new location

Run the Generated MEX with Updated Learnables

Call predict on the input image and compare the results with MATLAB.

clear mLayer_mex;
outNew = mLayer_mex(im,cnnMatFile)
outNew_MATLAB = mLayer(im,cnnMatFile)

outNew = 

  6(C) x 1(B) single dlarray

    0.1408
   -0.0080
    0.0342
   -0.0065
    0.1843
    0.0799


outNew_MATLAB = 

  6(C) x 1(B) single dlarray

    0.1408
   -0.0080
    0.0342
   -0.0065
    0.1843
    0.0799

Limitations

Only the network learnables and states can be updated by using the coder.regenerateDeepLearningParameters function. For modifications that the code generator does not support, an error message is thrown. For example, using coder.regenerateDeepLearningParameters after changing the scale factor of a leaky ReLU layer throws the following error message as scale factor is not a network learnable.

Network architecture has been modified since the last code generation. Unable 
to accommodate the provided network in the generated code. Regenerate code 
for the provided network to reflect changes in the network. For more 
information, see Limitations to Regenerating Network Parameters After Code Generation.